On Thu, Sep 17, 2026 at 04:59:37PM +0200, Jörg Rödel wrote: > Hi Tony, > > On Thu, Sep 17, 2026 at 09:27:47AM +0300, Tony Lindgren wrote: > > Oh OK, so calling for both hypervisor and confidential VCPU state would be > > needed. The difference could be specified in the struct kvm_vcpu_transfer > > flags? > > Or hypervisor state is migrated outside of the CVM live migration interface? > From QEMUs > perspective it is not different to other state like for emulated devices and > such. Does TDX also have per-cpu hypervisor state that needs to be migrated? > > In any case, for SNP this state must be settable from QEMU to correctly > support > IGVM loading, so an interface outside of live migration is needed :)
Oh OK so the per-cpu hypervisor state is part of the IGVM loading. Yes the per-vcpu migration is a bit different for TDX. No two separate steps like you're describing for vCPU migration (per-cpu for vCPU registers and per-cpu for hypervisor state). After stopping the TDX source, the VM state is migrated. And then the per-cpu state. The per-vcpu state is just a QEMU VMState for TDX too, and it needs the help of KVM to export the state into an encrypted blob.
